Memnest

Local-first memory for AI coding agents, via one MCP service

Editorially reviewedChecked Sep 2026How we review

mcp#mcp#ai-memory#local-first#coding-agents

Free plan Claimed API Self-hosted

Toolglade’s take

Memnest is a genuinely well-engineered take on a real problem: giving coding agents persistent, private memory without shipping your context to a third party. The local-first architecture (SQLite source of truth, rebuildable BM25 + vector indexes, an encrypted vault, and no external LLM calls) is careful and well-documented. The honest caveats are that it is very early-stage with minimal adoption so far, it is self-host only with a Linux-first install path, and it takes some manual setup to wire into each agent. Worth a look if you value privacy and self-hosting over one-click convenience.

About Memnest

Memnest is an open-source, local-first memory service for AI coding agents. A single Rust service stores durable memories and redacted conversation history on your machine and serves them to pi, Claude Code, Codex, and other MCP clients over MCP/HTTP. Retrieval uses local BM25 + vector hybrid search with no external LLM calls, and secrets live in an encrypted vault. It is self-hosted, MIT-licensed, and free.

Memnest solves the fact that AI coding agents forget everything when a session ends. It runs as one Rust service that captures durable memories (decisions, corrections, and preferences you save on purpose) plus redacted conversation history, and exposes them to any MCP client over a local endpoint (http://127.0.0.1:3111/mcp) or a plain HTTP API. Retrieval is a local hybrid search: BM25 keyword matching and HNSW vector similarity over locally-run multilingual embeddings, fused with RRF and reranked with MMR, so both exact tokens (a port number, a crate name) and paraphrases are found without any cloud calls. SQLite is the source of truth; the text and vector indexes beside it are derived and rebuildable. Memory is workspace-isolated, and credentials live in a separate AES-256-GCM encrypted vault rather than in searchable memory. It integrates with pi (via the pi-memnest npm extension), Claude Code and Codex (via a prompt hook and MCP), and any other MCP client. The trade-offs are that it is self-hosted only (no managed option), currently Linux-first for prebuilt installs (Windows/WSL via scripts, or build from source), and very early-stage as a project. But for developers who want private, on-machine agent memory with no vendor lock-in, it is a thoughtfully engineered option.

Weighing your options?See how Memnest compares to the alternatives.

TL;DR

Memnest is an open-source, local-first memory service for AI coding agents. One Rust service stores durable memories and conversation history on your machine and serves them to Claude Code, Codex, pi, and other MCP clients over MCP/HTTP, using local hybrid (BM25 + vector) search with no external LLM calls. It is self-hosted, MIT-licensed, and free, with an encrypted vault for secrets. Well-engineered but very early-stage, and Linux-first for prebuilt installs.

Company overview

Memnest is an independent open-source project by developer Blue-B, released under the MIT license. It is not a company or commercial product; it is a community-maintained tool distributed via GitHub with a companion npm package (pi-memnest).

The project focuses narrowly on private, local-first memory for AI coding agents and is at an early stage of development with a small user base as of 2026.

Product features

Core capabilities include durable memory storage, redacted conversation-history capture, and local hybrid retrieval combining BM25 keyword search and HNSW vector similarity (fused with RRF, reranked with MMR). SQLite is the canonical store; the keyword and vector indexes are derived and rebuildable.

Additional features: workspace isolation with a shared 'playbook' scope, an AES-256-GCM encrypted secret vault, prompt-time recall via a hook, transcript capture via a watcher, and both an MCP endpoint and a plain HTTP API. Embeddings run locally (multilingual e5-base) with no external LLM calls.

Target market

Individual developers using MCP-compatible coding agents (Claude Code, Codex, pi) who want persistent, private, self-hosted memory rather than a cloud service.

Buyer personas

End users

Developers who use AI coding agents and want them to remember context across sessions.

Buyers

Same as end users — it is free and self-installed, so there is no purchase.

Key influencers

Open-source and MCP-ecosystem communities and privacy-focused developers.

Ideal customer profile

A technical, privacy-conscious developer comfortable self-hosting a local service and wiring it into their MCP-based coding workflow.

Funding & performance

None disclosed. Memnest is an independent, MIT-licensed open-source project rather than a funded company; verify status on GitHub.

Pros & cons

Pros

Fully local-first — no data leaves your machine, no external LLM calls
Works across pi, Claude Code, Codex, and any MCP client via one service
Hybrid BM25 + vector search finds both exact tokens and paraphrases
SQLite source of truth with rebuildable indexes (resilient to corruption)
Encrypted AES-256-GCM vault for secrets, kept separate from memory
Open-source (MIT) and free, with no vendor lock-in
Thoughtful, well-documented architecture and security model

Cons

Very early-stage project with minimal adoption so far
Self-hosted only — no managed or cloud option
Prebuilt install is Linux-first (Windows/WSL via scripts or a source build)
Requires manual setup to wire into each agent/harness
No team collaboration or shared-server hosting out of the box
Small single-maintainer project; longevity is unproven

Pricing plans

Open Source (MIT)

Free

Self-hosted Rust service
MCP + HTTP API
Local hybrid (BM25 + vector) search
Encrypted secret vault
No cloud, no external LLM calls

Key features

API

Self-hosted

Multi-language

Integrations

Claude Code, Codex, pi, MCP clients, HTTP API

Input types

text

Output types

text

Best For

Persistent agent memory, Local-first / private setups, MCP-based coding workflows, Self-hosting developers

Compare key features

View all alternatives →

Feature

Memnest

ACI.dev

mcp-agent

Pricing

Free

Freemium

Free

Free plan

Yes

Free trial

API

Yes

Self-hosted

Yes

Team support

Yes

Frequently asked questions

Is Memnest free?+

Yes. Memnest is fully open-source under the MIT license and free to use. There is no paid or cloud tier; you self-host the service on your own machine.

Which agents does it work with?+

It works with pi (via the pi-memnest extension), Claude Code and Codex (via a prompt hook plus MCP), and any other MCP client pointed at its local /mcp endpoint. A plain HTTP API is also available.

Does Memnest send my data anywhere?+

No. It is local-first: memories, indexes, and embeddings all live on your machine, and it makes no external LLM calls. The server binds to localhost by default.

How does its search work?+

It combines BM25 keyword matching and HNSW vector similarity over locally-run multilingual embeddings, fused with RRF and reranked with MMR, so both exact tokens and paraphrases are retrieved.

Can I run it on Windows?+

Yes, via the provided Windows/WSL install scripts or by building from source with a Rust toolchain. Prebuilt one-command installs are currently Linux-first.